Meta

Key Terms

• Muse Spark - Meta's proprietary (closed-weight) AI model family, developed by Meta Superintelligence Labs (MSL) under Alexandr Wang. First announced April 8, 2026 as Muse Spark 1.0 (consumer-only, powering Meta AI chat). Version 1.1 launched July 9, 2026 with developer API access. Not open-weight. Source: Meta – Introducing Muse Spark Msl
• Meta Model API - Meta's self-serve hosted inference API at api.meta.ai/v1, in public preview for US developers as of July 2026. OpenAI SDK-compatible and Anthropic Messages-compatible. Provides $20 in free credits at signup. Supersedes the former Llama API waitlist (llama.developer.meta.com). Source: Meta – Build With Muse Spark
• Reasoning model - a model that performs internal chain-of-thought before answering. Muse Spark 1.1 is a reasoning model with configurable reasoning_effort levels (minimal to xhigh). Reasoning tokens are billed as output tokens and reported in usage.completion_tokens_details.reasoning_tokens. Source: Meta – Build With Muse Spark
• Prompt caching - repeated context in API requests is cached and billed at a discounted rate. Muse Spark 1.1 charges $0.15 per MTok for cached input, compared to $1.25 for fresh input, an 88% discount. Source: Dev – Pricing Rate Limits
• Open-weight model - model weights are published for download, allowing anyone to run inference, fine-tune, or modify the model locally. Llama 4 Scout and Maverick are open-weight under the Llama 4 Community License. Muse Spark is NOT open-weight. Source: GitHub – License
• Mixture-of-experts (MoE) - architecture where only a subset of parameters are activated per token. Llama 4 always activates 17B parameters per forward pass but routes through different expert subsets (16 for Scout, 128 for Maverick). Source: Openrouter – Llama 4 Maverick
• Llama 4 Community License - Meta's custom license for Llama 4 models. Permits free commercial use with two conditions: (1) entities with more than 700M monthly active users must request a separate license, and (2) derivative models must prefix their name with "Llama" and display "Built with Llama." Source: GitHub – License
• Computer use - an agentic capability where the model operates a GUI by reading screenshots and sending mouse/keyboard actions. Muse Spark 1.1 supports computer use in a sandboxed Linux desktop. Source: Meta – Build With Muse Spark

Latest Changes

Changes since the 2026-06 report.

Verification of June watch items:

• Llama 4.1 / Llama 5 / Behemoth announcement? FAILED. No Llama successor was announced. Instead, Meta pivoted to the Muse Spark product line. Llama 4 Scout and Maverick (April 2025) remain the only open-weight flagships, now 15+ months old. Source: Meta – Llama 4
• Llama API exiting waitlist? CONFIRMED (with twist). The Llama API waitlist at llama.developer.meta.com has been superseded by the Meta Model API at dev.meta.ai / api.meta.ai/v1, now in public preview for US developers. However, the Meta Model API serves Muse Spark, not Llama models. Llama weights remain download-only. Source: Meta – Build With Muse Spark
• Any new model releases? CONFIRMED. Muse Spark 1.1 launched July 9, 2026 (blog post July 8), available via Meta Model API and OpenRouter (July 16). Source: Meta – Introducing Muse Spark Meta Model Api
• Third-party pricing changes (OpenRouter, etc.)? CONFIRMED. Maverick input rose from $0.15 to $0.20 per MTok (33% increase), output from $0.60 to $0.80 per MTok (33% increase). Scout price unchanged at $0.10/$0.30 but context window on OpenRouter reduced from 10M to 1M in metadata. Source: Openrouter – Llama 4 Maverick

New items for July 2026:

• [NEW CATEGORY] Muse Spark 1.1 launched. Meta released Muse Spark 1.1 on July 9, 2026, its first proprietary (closed-weight) model available to developers via API. Muse Spark is a multimodal reasoning model with 1M context, supporting text, images, video, audio, and PDF input. It was originally announced as Muse Spark 1.0 on April 8, 2026 (consumer-only, powering Meta AI). Version 1.1 is the first developer-accessible release. Source: Meta – Introducing Muse Spark Meta Model Api
• Meta Model API in public preview. Self-serve API at api.meta.ai/v1, OpenAI SDK-compatible, with $20 in free credits for new accounts. US-only. Supports structured output, parallel function calling, built-in web search with citations, configurable reasoning effort, streaming, and prompt caching. Also supports Anthropic Messages format for Claude Code integration. Source: Meta – Build With Muse Spark
• Muse Spark 1.1 available on OpenRouter. Listed at $1.25/$4.25 per MTok since July 16, 2026. US-only. Single provider (Meta-hosted). Source: Openrouter – Muse Spark 1.1
• opencode built-in Meta provider. opencode ships with a built-in Meta provider: install, get a key from the Model API dashboard, run /connect, filter to Meta, paste key, select Muse Spark 1.1. Source: Meta – Build With Muse Spark
• Maverick price increase on OpenRouter. Llama 4 Maverick rose from $0.15/$0.60 to $0.20/$0.80 per MTok (33% increase on both input and output) between June and July 2026. Source: Openrouter – Llama 4 Maverick
• Scout context window reduced on OpenRouter. OpenRouter metadata now lists Scout context as 1M, down from the 10M listed in June. The model natively supports 10M context per Meta's model page, suggesting OpenRouter is capping served context. Source: Openrouter – Llama 4 Scout ; Meta – Llama 4
• Artificial Analysis Intelligence Index: 51. Muse Spark 1.1 scored 51 on the AA Intelligence Index (up from 43 for Muse Spark 1.0), effectively tied with GLM-5.2, GPT-5.4, and GPT-5.6 Luna. Coding Index: 71 (up from 59). SciCode: 58%, ranking #3 across all benchmarked models. Source: Artificialanalysis – Muse Spark 1 1 Everything You Need To Know
• Terminal-Bench 2.1 controversy. An HN commenter (self-identified ex-Meta employee) flagged that Meta's Terminal-Bench 2.1 evaluation used 6 CPU cores and 8GB RAM, exceeding the per-task resource caps defined in the benchmark (most tasks allow 1-4 CPU cores, many at 2GB RAM). Muse Spark 1.1 is not on the official Terminal-Bench leaderboard. Source: News – Item
• Strategic pivot from open-weight to proprietary. The New Stack reported in April 2026 that "Meta abandons open-source Llama for proprietary Muse Spark." The developer site (developer.meta.com/ai/) now lists Muse Spark as the primary model, with Llama 4 and Llama 3 demoted to secondary navigation links. Source: Thenewstack – Meta Abandons Llama Spark
• Behemoth still MIA. Llama 4 Behemoth, mentioned at the April 2025 launch as a larger model in training, has had no status update in over 15 months and appears effectively cancelled given the Muse Spark pivot. Source: Meta – Llama 4
• Benchmark credibility still unrebutted. The January 2026 report quoting departing Meta AI leadership confirming Llama 4 benchmark numbers were "fudged a little bit" remains unaddressed by Meta. Source: Tech – Results Were Fudged Departing Meta Ai Chief Confirms Llama 4 Benchmark Manipulation

Plans

Meta offers two access paths. The Meta Model API (for Muse Spark 1.1) is the primary developer product. Llama 4 models remain free to download under the Llama 4 Community License.

API Pricing

Meta Model API (Muse Spark 1.1)

Meta's first-party API, in public preview for US developers. All prices per 1M tokens.

Source: Meta – Build With Muse Spark ; Dev – Pricing Rate Limits

Third-Party API Pricing (Llama 4 and Muse Spark 1.1)

Performance characteristics (Muse Spark 1.1, Meta first-party API)

• Output speed: ~114 tokens/s median. Source: Artificialanalysis – Muse Spark 1 1 Everything You Need To Know
• Time to first answer token: ~21s. Source: Artificialanalysis – Muse Spark 1 1 Everything You Need To Know

Self-hosting hardware requirements (Llama 4)

• Llama 4 Scout: Can run on a single H100 GPU with INT4 quantization (109B total params). Source: Meta – Llama 4
• Llama 4 Maverick: Cannot run on a single GPU. FP8 quantized weights fit on a single H100 DGX host (8 GPUs). BF16 weights require multi-host deployment (400B total params). Source: Hugging Face – Llama 4 Maverick 17B 128E

Model Performance / Benchmarks

Muse Spark 1.1 (Meta-reported)

Meta's published benchmarks for Muse Spark 1.1, compared against frontier competitors. Note: Terminal-Bench 2.1 results have been challenged for resource-limit violations (see Terminal-Bench controversy in Latest Changes and Community Signals).

Source: Meta – Muse Spark ; Meta – Muse Spark 1 1 Evaluation Report

Muse Spark 1.1 (Artificial Analysis, independent)

Source: Artificialanalysis – Muse Spark 1 1 Everything You Need To Know

Llama 4 (Meta-reported, legacy)

These scores are from the original April 2025 launch. Note: departing Meta AI leadership confirmed in January 2026 that Llama 4 benchmark numbers were "fudged a little bit." Source: Tech – Results Were Fudged Departing Meta Ai Chief Confirms Llama 4 Benchmark Manipulation

Source: Meta – Llama 4

Latest News

Muse Spark 1.1 and Meta Model API Launch (July 9, 2026)

Meta launched Muse Spark 1.1 and the Meta Model API in public preview for US developers. The blog post, published July 8, describes Muse Spark 1.1 as a "step-change" over the April 2026 version 1.0, with improvements in agentic workflows, coding, and multimodal perception. The API is OpenAI SDK-compatible, supports Anthropic Messages format, and ships with built-in web search grounding, streaming, and prompt caching. opencode has a built-in Meta provider. The announcement generated significant attention: 413 points and 214 comments on HN. Bloomberg reported Zuckerberg is pledging "aggressive" pricing, framing this as Meta's "first pay-to-use AI." Source: Meta – Introducing Muse Spark Meta Model Api ; Meta – Build With Muse Spark ; Archive – 3Ccka

Strategic Pivot from Open-Weight Llama to Proprietary Muse Spark

Muse Spark is Meta's first proprietary (closed-weight) model. The developer site now lists Muse Spark as the primary product, with Llama 4 and Llama 3 relegated to secondary navigation. The New Stack reported in April 2026 that this represents Meta "abandoning open-source Llama for proprietary Muse Spark." Alexandr Wang, who leads Meta Superintelligence Labs, previously made his money selling data annotation services to closed-source AI companies including OpenAI, Anthropic, Google, and Meta. The open-weight community has noted the shift, with HN commenters observing that "the muse family is no longer open-source." Sources: Thenewstack – Meta Abandons Llama Spark ; Axios – Meta Muse Alexandr Wang ; News – Item

Artificial Analysis: Muse Spark 1.1 Scores 51 on Intelligence Index (July 10, 2026)

Independent evaluator Artificial Analysis reported Muse Spark 1.1 at 51 on their Intelligence Index, up 8 points from Muse Spark 1.0 (43) in three months. The model is effectively tied with GLM-5.2, GPT-5.4, and GPT-5.6 Luna, and is described as "the most token-efficient of the models effectively tied at 51." SciCode at 58% ranks #3 across all models benchmarked. Cost per Intelligence Index Task is estimated at ~$0.26. Source: Artificialanalysis – Muse Spark 1 1 Everything You Need To Know

Build-Off Comparison: Muse Spark vs GPT-5.6, Grok 4.5, Claude (July 10, 2026)

A multi-model build-off comparing 12 models building the same 4 apps (including Muse Spark) reached 159 points and 89 comments on HN. The comparison provides practical coding-quality data beyond benchmark scores. Source: Tryai – Gpt 5.6 Build Off 12 Models ; News – Item

Llama 4 Models Still Available but Superseded

Llama 4 Scout and Maverick remain downloadable from developer.meta.com and are still served on OpenRouter, but they are clearly positioned as legacy products. No Llama 4.1, Llama 5, or Behemoth has been announced. The old Llama API waitlist has been replaced by the Meta Model API, which serves Muse Spark exclusively. Source: Meta – Llama 4

Benchmark Manipulation Still Unrebutted (January 2026, ongoing)

The January 2026 report quoting departing Meta AI leadership confirming that Llama 4 benchmarks were "fudged a little bit" remains unaddressed. Meta has not restated scores or published corrections. This history is relevant context for evaluating Muse Spark 1.1's published benchmarks, particularly given the Terminal-Bench 2.1 resource-limit controversy. Source: Tech – Results Were Fudged Departing Meta Ai Chief Confirms Llama 4 Benchmark Manipulation

Community Signals

The Muse Spark 1.1 launch generated the largest HN discussion of any Meta AI product since the original Muse Spark announcement in April 2026 (393 points, 367 comments). The July 1.1 thread reached 413 points and 214 comments, with discussion focused on pricing competitiveness, benchmark integrity, and the open-weight-to-proprietary pivot.

Pricing competitiveness

Tiberium (HN): "The pricing is insane: $1.25/$4.5 for 1M tokens, and $0.15 for cached input!" News – Item

ignoramous (HN): "Cheaper than Qwen 3.7 Max. Second indication, after Grok 4.5 ($2 in / $6 out), that the BigLabs are feeling the GLM 5.2 heat." News – Item

mchusma (HN): "this is most directly comparable to xAI Grok 4.5. In both cases, directionally 'opus level intelligence for haiku prices' which is a really big deal for application developers who want to include models like this in their applications." News – Item

Aurornis (HN): "The cached input pricing is a good ratio. Compare with Grok 4.5 which came out at $2/$6 but then quietly charges $0.50 per 1M cached input tokens. That's as high as Opus 4.8!" News – Item

Benchmark integrity controversy

GodelNumbering (HN, self-identified ex-Meta employee): "From Terminal-bench-2.1 details: 'We use a bash-tool-only agent harness to evaluate 89 Terminal-Bench 2.1 tasks from the official repository, where resources are capped at 6 CPU cores and 8GB RAM.' This disqualifies the results. Each terminal bench task has a cpu upper limit and RAM upper limit. Overriding either is disqualification. [...] This kind of shady benchmarking takes all joy out of building a harness to improve benchmark performance of a model because no matter what you do, you won't beat the headline (cheating) number. This is presumably why this model is not in the official benchmark leaderboard. As an ex Meta employee, this is a little sad but not massively surprising. 'Number go up' is the core performance evaluation metric until PSC is done and you move on." News – Item

Open-weight-to-proprietary pivot

rldjbpin (HN): "the muse family is no longer open-source. it is still priced very 'cheaply', but a significant shift in approach." News – Item

jacobgold (HN): "Maybe Zuck should double down on his 'spoiler' role with models rather than compete head-to-head. He doesn't have to match Anthropic or OpenAI model revenue if he can deflate theirs by 99%. All he has to do is keep spending a few billion dollars developing frontier models, release them as open weights, and turn coding models into a commodity." News – Item

alansaber (HN): "Meta's local llama models used to be the face of open source AI. The scene has really changed." News – Item

General reception

kilroy123 (HN): "I personally do not like Meta, but I'll say this. The more competition, the better for regular consumers. I'm building like crazy to take advantage of all these subsidized tokens while I can." News – Item

simonw (HN, Simon Willison): "I had a few days of preview access, which was long enough to put together a plugin for LLM. You can try the model out in the terminal like this: llm -m meta-ai/muse-spark-1.1 'Generate an SVG of a pelican riding a bicycle'" News – Item

Industry quotes (from Meta's launch blog)

Amjad Masad (CEO, Replit): "What's most impressive about Muse Spark is how much it packs into one model: massive million-token context, full multimodal support (images, video, PDFs), built-in search with citations, strong reasoning, top-tier coding abilities (particularly frontend and design), structured output, and parallel tool calling, all in a clean OpenAI-compatible package." Source: Meta – Build With Muse Spark

Saoud Rizwan (CEO, Cline): "Meta is clearly building for serious agentic coding, strong tool use at a price point that makes it viable to run real coding workloads at scale. That combination is rare, and it's exactly why we wanted Cline developers to have access early." Source: Meta – Build With Muse Spark

Enterprise Readiness

Source: Meta – Build With Muse Spark ; GitHub – License

Transparency Gaps

---

HLD update recommendation: The HLD.md entry for Meta should be updated to reflect the new flagship model and URLs:

• Latest LLM: Muse Spark 1.1 (primary), Llama 4 Scout, Llama 4 Maverick (open-weight legacy)
• Pricing URL: https://developer.meta.com/ai/ (llama.com now redirects here)
• Changelog / Blog URL: https://developer.meta.com/ai/resources/blog/

This requires user approval per the data separation rules. Muse Spark is not a new supplier (same Meta entity), but it is a new product category (proprietary API vs. open-weight download).